In a physical activity lifestyle program aimed at promoting a healthier lifestyle consisting of more physical exercise, a mobile body-worn activity monitor with a built-in triaxial accelerometer was provided to participants to record daily physical activity level (PAL) data. The PAL data quantifies the level of a person's daily physical activity and reflects the daily energy expenditure of this person. In the current lifestyle program, coaches help and motivate participants to achieve a pre-set increased level of physical activity within 12 weeks. However, many participants of such programs fail to complete the program; they drop out. As a result, they do not attain the desired increase in physical activity. This thesis researches methods to predict PAL from past data and to predict early dropping out of the program in a classification task. By detecting participants who are at risk of dropping out of the program in advance, it helps to provide a timely delivery of interventions and motivating triggers. In other words, it will enlarge the reach of coaches by a better understanding of the participants in the program. In particular, this thesis proposes a categorized autoregressive integrated moving average (CARIMA) method to predict future PAL data of every next week from past data of participants. It achieves a large reduction in computation time with little loss in prediction accuracy compared with the traditional ARIMA models. The results show that our CARIMA method performed well in terms of prediction accuracy (~9% mean absolute percentage error), model parsimony and robustness. For the dropout prediction, we combined our set of features from the ARIMA process with those features extracted from an activity database containing participant characteristics and activity data to classify dropouts of every next week. A genetic algorithm was applied to combine and select features and principal components analysis was used to reduce feature vector dimension. The unbalanced data set problem was handled by using a weight cost sensitive learning method. Results show that a k-nearest- neighbor classification algorithm achieved an average dropout and non-dropout classification accuracies of ~66% and ~74%, respectively.
|Date of Award||31 Aug 2009|
|Supervisor||Ronald M. Aarts (Supervisor 1)|